Experience

Work Experience

Conducted experiments with zero-shot and few-shot prompt engineering techniques initially on ChatGPT, then on open-source Large Language Models (LLMs) like LLaMA to create an unstructured to structured requirement converter.
Designed a multi‑panel visualization tool utilizing Flask and React, Docker, and MongoDB to convert semi‑structured requirements into knowledge graphs and structured formats.
Built an active learning pipeline to facilitate the continuous training and testing of the CodeT5 model.
Devised a GPU enabled Docker containerized LLM API and CodeT5 converter API using LangChain.
Currently fine-tuning Mistral-7B with custom unstructured to structured requirement conversion data using PEFT and LoRA

Evaluated the performance of BERT based state-of-the-art Natural Language Processing (NLP) models like ALBERT, RoBERTa, and DistilBERT for automatic data preprocessing and annotation at the research and development phase.
Fine‑tuned an Electra-QA model to annotate watch features from raw descriptions.
Over a 10‑fold reduction in data annotation costs and time by leveraging large language models for data annotation.
Research Engineer | United International University | Dhaka, Bangladesh | March 2022 - September 2022

Developed full-stack artificial intelligence based Smart Receptionist web application, with React frontend, Flask backend, and MySQL database.
Deployed the microservices on a Linux server using Docker and Nginx containerization.

Built NLU, an intent and named entity recognition (NER) engine, capable of integrating into existing chatbots with FastAPI.
crafted a complex JSON parser to store NLU data into MySQL database.
Devised an AI voice assistant based on the NLU engine.

Created computer vision based REST API using OpenCV and TensorFlow for age-gender prediction, and emotion recognition from images and videos.
Evaluated YOLO object detection algorithm for inference and memory usage.
Experimented with TFLite and ONNX conversion of object detection models.
Enhanced deep learning based object detection inference time by over 50% by resolving CUDA GPU memory leaks in Darknet and deployed containerized microservice on AWS EC2.
Composed a scheduled batch upload of object detection results to AWS S3 using Celery and Redis

Created a product search engine using Elasticsearch with text and categorical filters.
Integrated image recommendations to existing recommendation engine through image processing and feature vectorization.

Implemented a machine learning based budget prediction feature for online ad campaigns.

Configured a private Instant Messaging (IM) server to facilitate peer-to-peer (P2P) connections for audio, video, and text messaging.
Used prosody as the XMPP communication server and Jitsi as the XMPP client.